The Cycle of Human-AI-Human in Nudging

To readers who know Richard Thaler and Cass Sunstein’s original Nudge (2008), it might come as a surprise that they have named the second edition of the book Nudge: The Final Edition. “Final” is a claim so bold even for these Nobel-Prize winning authors that even they joke about it. In Thaler and Sunstein’s defense, they have all the right they want to call this one by such a title. With 13 years between two books, the authors were able to rethink some old problems mentioned back in the first edition of the book, and discover some new problems and ideas not wrote about in the first edition, and Nudge: The Final Edition covers the new understandings by Thaler and Sunstein on the topic of nudging and Libertarian Paternalism. Some are problems which seemed complicated back then are solved now, and there are some new problems that arise during these 13 years which are also addressed in this edition. But no core concept has changed. If anything, the passage of time has clarified this theory and its applications.

The book’s theory about nudging can be summed up in one sentence. “How choices are presented and what choices are presented affect how people choose.” Then after defining what nudge is and recognizing its place in our lives, the book went on to introduce what they call “Libertarian Paternalism”, the ideal of nudging people to make good choices about their lives that sometimes may not be so obvious or otherwise they have no incentive in choosing. The book advocates for this ideal as they think nudging is a way of motivating people without interfering with their freedom of will. This is a controversial view that the authors are optimistic about, as some other people believe since nudging is so powerful, people could hardly keep their choices to themselves and choice architects are more likely to exploit nudging for benefits rather for people’s welfare. No matter where nudging eventually ends up at, we should not overlook this powerful theory that explain how all choices presented to us in this world affect us. At the core of nudge is choice architecture, how choices are constructed. Everyone who are presenting choices for others are as choice architects who can deploy numerous tools to nudge people in preferred directions. Thaler and Sunstein listed some essential aspects to consider when constructing a choice architecture: default choice, expected error, feedback, choice to welfare mapping, complex choices, incentives, curation, and fun. Some names are quite representative of the categories. Default choice are the ones you see on websites with cookie settings where they checked the “agree” box already for you, feedback is what the choice maker get after they made a choice, incentives are what prompts people to make a specific choice, and fun is simply fun, what people can get from the choices they made. Some are more abstract: expected error are situations where the choice architect would leave some room for people to make mistakes but will still cover them, complex choices talks about a problem where if an architect is too complex/there are too many choices, people are less inclined to think before they act, such as some website’s overly complicated cookie statement that user need to go through before they say no to everything. The other are less important as they point to very specific field.

From the examples derived from the book and the toolbox explained, it is not hard to see that nudging is deeply embedded into this world, into every single situation where choices are presented and people need to make a decision. The tools are surprisingly simple as well. They are not novel things that no one has thought about or no one has created before. The methods mentioned are simply summarization of what people have been doing for a long time either intentionally or unintentionally to nudge others into one direction. This would support a point made by the authors of the book — nudge is unavoidable, and you cannot get rid of it. That is why the authors are advocating for Libertarian Paternalism, a way to nudge people for good. However just like with any other technique, theory, methodology, technology… whatever novel development there maybe, once people have found something else that could be linked to this pre-existing thing, they would try to link them together and take one step forward for better or worse. In this case, it is a field overlooked by Thaler and Sunstein: nudging by AI which only became available in recent years with gasping advancements in related fields of research, and even more so with a disaster to force everyone online, the COVID-19 pandemic.

With COVID-19 hitting the world like a train wreck and people have been living virtually for an extended period of time, Thaler and Sunstein has put some attention on how internet, advertisement algorithms and etc play into nudging. But they never went into depth regarding this part. There is a deeper dive that can be taken in this direction. It is a consensus at this point that AI is everywhere in our lives and have been nudging people left and right based on intentions of their creators. But it is less known to the public that, Machine Learning, the “autonomous and individual” learning process which produces AI can be affected by human biases. This therefore creates a loop where human nudges AI, then AI nudges human, and the nudged human nudges AI again. While some people say this is a malicious cycle, it doesn’t necessarily have to be. Both human bias and AI’s lack of ethic can be addressed in this process, and by a single change.

To understand how nudging can go both ways in context if AI-Human relationship, it is important to understand how AI works. AI as people generally refers to is the result model of Machine Learning (aka ML), where quite literally, a machine learns like a human to solve a problem that is given to it. For nudging there are three important pieces of ML that we need to talk about: training dataset, validation dataset, and learning algorithm. Training and validation datasets can be explained by comparing them to practice problems and actual exam questions for students. Human students practice example problems to understand concept and explore ways to solve the problem, and they are then tested with exam questions to see how well they perform and if there is need to change their learning approach to the problem. Then the algorithm can be said as a learning method. How students learn, which for different students there are different methods: some read others’s solutions extensively before attempting a single problem on their own, while some dive head first and only refer to other’s solution when they are stuck. And from this explanation, it is easy to see how AI can be nudged by human. We can either give it biased problems sets, or we can give it a modified algorithm.

Human bias’s affect can be seen clearly in an example described in an article published by Propublica regarding a risk assessment AI for criminals. With the only human input being the dataset it was given, which is a list of criminals informations and whether they committed crime again after serving sentence for their first one, the AI became racist because it categorizes black criminals of being higher risk people with higher possibility to commit crime again. The problem actually is with the dataset it was given, due to human bias, the dataset included more examples of black criminals committing crimes once more after they have served sentence for the first crime. (Propublica)

Just talking about how human can nudge AI is not enough, not enough for real world applications, not enough for context of this book. The book talks about a very important concept of Libertarian Paternalism, the idealism of using nudging to motivate people to make better choices for themselves. However this concept is very humanistic due to it being based on the fact human knows what is morally correct and what is morally wrong. AI on the other hand has no idea what moral is, and therefore knows no right or wrong in that context. Propublica’s example showcases a problem solving AI without input on moral values can be a good problem solver that is morally corrupted. However this is not a unsolvable problem, which is very interestingly showed by a purposely constructed morally corrupted AI: introducing GPT-4chan by Kilcher.

Kilcher, a skilled ML researcher, one day came up with the brilliant idea to train a chatbot using 4chan’s infamous /pol/ subsection. For those unfamiliar with 4chan/pol/, the pol here does not stand for politics, but rather politically incorrect. Wikipedia summarizes this specific section being a discussion board for far-right extremists, with topics mostly centered around racism, white supremacism, anti-everything and etc. (Wikipedia) So yeah, you can say Kilcher purposefully told the AI to do the wrong thing. And that is what exactly came out of the ML process: an AI as degenerate as the members in that forum, with what it said being only more extreme than human posts on that discussion board. (Kilcher)

While GPT-4chan is a degenerated AI that should not be put to use anywhere, it showed how we can nudge AI in a positive direction, and in turn use AI to nudge people more effectively to achieve Libertarian Paternalism. It is clear at the current stage that AI exists everywhere in our lives and are nudging us all the time. This is best showcased with micro-targeting advertisement that shows up everywhere. While we are browsing websites we can see advertisement tailored to our interest, in our mailboxes and email inboxes there are auto-generated tailor-fitted coupons for us, all of which are based on our purchase history and browsing/searching history. AI learns them, and they could derive from that what we want and therefore tailor these things to nudge us to buy them. This is only one aspect of our lives that is heavily infiltrated by AI, there exist many other more. If we can even replace part of them with good-intention AI, we could achieve Libertarian Paternalism. For example, instead of nudging consumers to buy a certain product that makes the most benefit for the companies, nudge them to buy healthier products based on their health history. Give them suggestion on gym workouts, generate link to articles on how to obtain a healthier lifestyle. All of these are possible, and we just need to be determined to do them.

Just to wrap things all up, Thaler and Sunstein worked brilliantly to write this book: “Nudge: The Final Edition”. It again emphasized the points regarding nudge being everywhere and claim that this is beneficial to the world under libertarian paternalism. However as the most powerful form of nudging has became AI nudging, Thaler and Sunstein did not really give enough room for this topic.